FastForward Pruning: Efficient LLM Pruning via Single-Step Reinforcement Learning

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Yuan, Xin, Li, Siqi, Wei, Jiateng, Zhu, Chengrui, Wu, Yanming, Li, Qingpeng, Lv, Jiajun, Lan, Xiaoke, Chen, Jun, Liu, Yong
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866912726572335104
author Yuan, Xin
Li, Siqi
Wei, Jiateng
Zhu, Chengrui
Wu, Yanming
Li, Qingpeng
Lv, Jiajun
Lan, Xiaoke
Chen, Jun
Liu, Yong
author_facet Yuan, Xin
Li, Siqi
Wei, Jiateng
Zhu, Chengrui
Wu, Yanming
Li, Qingpeng
Lv, Jiajun
Lan, Xiaoke
Chen, Jun
Liu, Yong
contents Pruning is an effective method for compressing Large Language Models, but finding an optimal, non-uniform layer-wise sparsity allocation remains a key challenge. While heuristic methods are fast but yield suboptimal performance, more powerful search-based approaches like Reinforcement Learning are often hindered by prohibitive computational costs on large-scale models. To overcome this efficiency barrier, we propose FastForward Pruning. Its core is a decoupled, single-step RL framework that separates policy optimization from the complex budget satisfaction problem. Such a decoupling is crucial for efficiently searching the vast policy space of LLMs. This curriculum-based strategy begins with low-cost, simple tasks and gradually increases in complexity, significantly reducing the search's computational overhead. Evaluated on the LLaMA, Mistral, and OPT model families, our framework discovers pruning policies that achieve superior performance over strong heuristic baselines. Crucially, when compared to other search-based algorithms, our method achieves competitive or superior results at a fraction of the computational cost, demonstrating a clear advantage in search efficiency.
format Preprint
id arxiv_https___arxiv_org_abs_2511_18977
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle FastForward Pruning: Efficient LLM Pruning via Single-Step Reinforcement Learning
Yuan, Xin
Li, Siqi
Wei, Jiateng
Zhu, Chengrui
Wu, Yanming
Li, Qingpeng
Lv, Jiajun
Lan, Xiaoke
Chen, Jun
Liu, Yong
Machine Learning
Artificial Intelligence
I.2.7; I.2.6
Pruning is an effective method for compressing Large Language Models, but finding an optimal, non-uniform layer-wise sparsity allocation remains a key challenge. While heuristic methods are fast but yield suboptimal performance, more powerful search-based approaches like Reinforcement Learning are often hindered by prohibitive computational costs on large-scale models. To overcome this efficiency barrier, we propose FastForward Pruning. Its core is a decoupled, single-step RL framework that separates policy optimization from the complex budget satisfaction problem. Such a decoupling is crucial for efficiently searching the vast policy space of LLMs. This curriculum-based strategy begins with low-cost, simple tasks and gradually increases in complexity, significantly reducing the search's computational overhead. Evaluated on the LLaMA, Mistral, and OPT model families, our framework discovers pruning policies that achieve superior performance over strong heuristic baselines. Crucially, when compared to other search-based algorithms, our method achieves competitive or superior results at a fraction of the computational cost, demonstrating a clear advantage in search efficiency.
title FastForward Pruning: Efficient LLM Pruning via Single-Step Reinforcement Learning
topic Machine Learning
Artificial Intelligence
I.2.7; I.2.6
url https://arxiv.org/abs/2511.18977